Replace lltm with myaddmul; update to new custom ops APIs

[zou3519]

Stack from ghstack (oldest at bottom):

• -> #94

We're replacing lltm with myaddmul(a: Tensor, b: Tensor, c: float), which just does a*b+c. This simplification allows us to focus on the operator registration instead of get lost in the details of the complicated lltm kernels.

Test Plan:

• tests

[zou3519]

Uh that's not how to land it